Parameter Retrieval of Transparent Cirrus Clouds over South China Sea Based on Artificial Neural Networks

Objective Cirrus clouds are located in the upper troposphere and lower stratosphere and are mostly composed of ice crystal particles with a cloud top pressure of less than 440 hPa. Meanwhile, they are widely distributed, covering an area of approximately 20%-35% of the earth. In tropical regions, the coverage area of cirrus clouds reaches 70%. Cirrus clouds play a very important role in the earth atmosphere system by reflecting solar shortwave radiation and absorbing surface thermal radiation. Passive remote sensing features wide spatial coverage and high temporal resolution, but its detection for thin cirrus clouds is relatively limited. A large number of thin cirrus clouds with optical depth of less than 0.4 are missed by passive remote sensing. This phenomenon is particularly serious in tropical areas, which greatly underestimates the coverage of cirrus clouds and causes large errors in radiation calculation. Fortunately, the dual-wavelength cloud aerosol lidar infrared with orthogonal polarization (CALIOP) onboard the CALIPSO has unparalleled advantages in detecting thin cirrus clouds, which can detect thin cirrus clouds with optical depth less than 2. However, CALIOP cannot penetrate thick cirrus clouds, with limited spatial coverage. Therefore, we combine active and passive remote sensing to identify and retrieve optical depths and top heights of single-layer transparent cirrus clouds over the South China Sea. We hope that our research can provide quantitative support for the distribution of thin cirrus clouds over the South China Sea, and help improve the radiation calculation accuracy. Methods Artificial neural networks have powerful nonlinear function fitting ability and can employ multi-channel information, which makes them widely applicable in cloud parameter retrieval. Based on a matched dataset from MODIS/CALIOP, we utilize back propagation (BP) neural networks for the identification and parameter retrieval of single-layer transparent cirrus clouds over the South China Sea. First, to obtain a perfect dataset, we conduct several steps, including MODIS/CALIOP data collection, data processing, data conversion, data resampling, data matching, and data filtering. Additionally, we collect MODIS/MYD02_1 km and CALIOP/Level2_CloudLayer_5 km data for a total of six years from 2009 to 2015 over the South China Sea. MODIS raw data are converted to brightness temperature as inputs of neural networks. Due to the spatial resolution of 1 km for MYD02 data, it is necessary to resample it to 5 km for data matching with CALIOP. Meanwhile, data filtering based on three conditions is conducted, including CAD_Score and Number_Layers_Found from CALIOP, and MODIS/CALIOP pixel distance. We set the threshold of CAD_Score as 50. Since we only focus on single-layer transparent cirrus clouds, Number_Layer_Found is set as 1 and the distance between MODIS/ CALIOP is ensured to be no more than one pixel (5 km). After filtering, a total of 274786 data samples were obtained from 2009 to 2015, including 120980 single-layer transparent cirrus samples and 153806 clear sky samples. To increase the proportion of positive samples (transparent cirrus clouds) to improve the model accuracy, we randomly downsample the clear sky samples by 0. 6 times to obtain 92284 clear sky samples. Therefore, a dataset containing 213264 samples is finally acquired. We divide the dataset into training, testing, and validation sets in a 6: 3: 1 ratio, and three neural networks are proposed, including one for detecting transparent cirrus clouds, one for retrieving optical depths, and one for retrieving top heights. Results and Discussions In the identification of transparent cirrus clouds, the probability threshold of the network output is set to 0. 56, which can achieve the optimal detection rate and false alarm rate, with a detection rate of 79%, a false alarm rate of 9. 8%, and an AUC of 0. 92 [Fig. 3(a)]. A large number of transparent cirrus clouds with optical depth less than 0. 1 result in a low detection rate and a high false alarm rate, which is because it is difficult to distinguish these cirrus clouds from clear skies. The detection rate rapidly increases with the rising optical depth [Fig. 3(b)]. When the optical depth is less than 0. 03 or greater than 0. 4, the detection rate is only 36% or over 95% respectively. Among them, when the optical depth is greater than 1, the detection rate can reach 100%, indicating that the neural network can detect transparent cirrus clouds in the region, but the detection performance on the subvisual cirrus clouds with optical depth less than 0. 03 is poor. For the optical depth retrieval of transparent cirrus clouds, the error exceeds 500% under the optical depth of less than 0. 03. As the optical depth gradually increases, the error rapidly decreases to within 100%. When the optical depth is greater than 0. 2, the error is within 50% [Fig. 5( a)]. The reason for the large overall error is that the neural network has a significant error in retrieving cirrus clouds with optical depth less than 0. 03. The error for the top height of cirrus clouds is mainly distributed around 5%. In the parts with cloud top heights less than 10 km and greater than 17. 5 km, greater errors can be observed [Fig. 5(b)], and the main reason for speculation is that the sample sizes in these two intervals are relatively small. The scatter plots reveal good linear relations between the predicted and true values ( Fig. 6), with correlation coefficients reaching 0. 79 (for optical depth) and 0. 87 (for top height). The mean absolute error and root mean square error for the optical depth are 0. 2 and 0. 25 respectively. The mean absolute error and root mean square error of the top height retrieval are 0. 61 km and 0. 74 km respectively. In a case study, the comparison with CALIOP data shows that the research results have a certain reliability degree (Figs. 9 and 10). Conclusions Based on the MODIS/CALIOP matched dataset, neural networks are adopted to first detect transparent cirrus clouds, then retrieve the optical depth and cloud top height of the detected transparent cirrus clouds, and compare the results with the CALIOP data. The results show that the classification neural network can detect transparent cirrus clouds with a detection rate of 79%. The retrieval results show that it has a high agreement with CALIOP data. The correlation of the optical depth is 0. 79 and that of the cloud top height is 0. 87. The classification neural network has poor detection ability for subvisual cirrus clouds with optical depth less than 0. 03, with a detection rate of only 36% and a larger retrieval error. As the optical depth increases, the detection rate rises rapidly. For the parts with optical thickness greater than 0. 4, the detection rate reaches more than 95% and the retrieval error decreases rapidly. Finally, we perform a case study. The results show that the detection results of the neural network are more consistent with the observations of CALIOP than the official MODIS product. The retrieval results are in good agreement with CALIOP data. The results can provide references for the distribution of transparent cirrus clouds missed by MODIS, and help improve the radiation calculation accuracy.